What it is
Command A+ is Cohere's newest Command model, released on 20 May 2026 as command-a-plus-05-2026. Cohere describes it as the last model in the Command A family and its first sparse mixture-of-experts model: 218 billion parameters, 25 billion active per token, drawn from 128 experts with 8 active plus one shared. It reads text and images and writes text, with 128K input tokens and up to 64K output tokens.
It supports 48 languages, including every official EU language, along with tool use with inline citations, switchable reasoning and translation. In its recommended W4A4 build, where only the experts are 4-bit and attention stays at full precision, it runs on one NVIDIA B200 or two H100s.
The weights are on Hugging Face under Apache 2.0, and the model is on Cohere's API, private deployments and OpenRouter. Cohere's own pricing page lists no per-token rate for it, and production rate limits are by arrangement with sales.
Our review
Command A+ is easy to approve and hard to evaluate. Cohere calls it the strongest agentic model in the Command family and claims up to 110% higher throughput and 30% lower latency than Command A Reasoning, but neither its docs nor its model card give a single benchmark score. For a model aimed at enterprise buyers, that leaves procurement relying on its own trials.
The deployment story is the strongest part. Self-hosted, Apache 2.0 carries no revenue cap or acceptable-use addendum; Cohere's usage policy applies only on its own deployments. The W4A4 build quantises the experts to 4-bit while keeping attention and the KV cache at full precision, which is how a 218B model fits on two H100s; FP8 needs four H100s and BF16 eight. Cohere says the three builds differ negligibly in quality but publishes no figures.
Price is the surprise. Cohere's pricing page lists only legacy Command models and sends Command A+ buyers to sales, yet OpenRouter carries it at $0.30 per million input tokens and $1.50 output, about an eighth of the input price it lists for the older Command A. On Cohere directly, trial keys are capped at 20 requests a minute and barred from commercial use, and production limits for Command A+ are by request.
Two gaps remain. The 128K input window is shorter than most current frontier rivals, and adoption is thin: the W4A4 repository logged 3,686 downloads last month, so community fine-tunes and serving recipes are scarce beyond Cohere's own vLLM path, which needs vLLM 0.25 and Cohere's cohere_melody library to parse responses correctly.
Where it earns its place
A platform team at an EU insurer with two H100s on-premises
Answer staff questions from policy documents in a dozen EU languages without data leaving the building
Serve the W4A4 build in vLLM 0.25 with citations enabled, so each answer links its spans to retrieved passages. Hardware is the only cost and Apache 2.0 needs no licence negotiation, but budget for your own evaluation, since Cohere publishes no scores.
A legal-tech startup building a contract-review agent
Run tool-calling agents over long contracts at low per-token cost
Through OpenRouter at $0.30 input and $1.50 output per million tokens, 100,000 reviews a month at 20,000 input and 2,000 output tokens is about $900. Contracts longer than 128K tokens must be chunked.
A procurement lead at a bank already using Cohere Embed and Rerank
Add a generative model under the same vendor contract and security review
Command A+ runs on Cohere's SOC 2 Type II platform or in a VPC, on-premises or Model Vault deployment, so it can sit under the existing agreement. Expect a custom quote; Cohere lists no per-token rate.
Before you commit
- Setup
- Download CohereLabs/command-a-plus-05-2026-w4a4 and serve it with vLLM 0.25 or later plus cohere_melody 0.9, or call command-a-plus-05-2026 on Cohere's API with a trial key.
- Integrations
- Cohere Chat API, OpenRouter, vLLM, SGLang, Transformers and Docker Model Runner; JSON-schema tool calling with citation tags; pairs with Cohere Embed and Rerank.
- Data and privacy
- Self-hosted weights keep data in-house under Apache 2.0. Cohere's API platform is SOC 2 Type II with configurable retention and a training opt-out; private deployments run in your VPC, on-premises or in Model Vault, where Cohere says it has no access.
- Cost at scale
- Self-hosting needs two H100s or one B200 in W4A4, four H100s in FP8 and eight in BF16. Hosted, 1M RAG answers a month at 8,000 input and 400 output tokens is about $3,000 at OpenRouter's $0.30/$1.50.
- Limits
- 128K input and 64K output tokens; trial keys 20 requests a minute and not for commercial use; production limits by request; no published benchmark scores.
What's good
- Apache 2.0 weights, no revenue cap
- W4A4 build runs on two H100s or one B200
- 48 languages, including every official EU language
- Inline citations in tool use for RAG
- OpenRouter lists $0.30/$1.50 per 1M tokens
What's not
- No published benchmark or quantisation-loss figures
- No per-token price on Cohere's own pricing page
- Production rate limits only through sales
- 128K input context
- Needs vLLM 0.25+ and Cohere's parser library
Best for
Enterprises building private, multilingual RAG and agent systems who want a permissively licensed model they can host on two GPUs.
Alternatives
Command A+: common questions
- Is Command A+ worth it?
- Apache 2.0 weights, two-H100 hosting and every EU language make Command A+ easy to approve, but Cohere still publishes no benchmark scores to judge it by. We rate it 3.8 out of 5.
- How much does Command A+ cost?
- Free weights under Apache 2.0. Cohere API rate not published; OpenRouter lists $0.30 input / $1.50 output per 1M tokens. Private deployment by quote.
- What are the downsides of Command A+?
- No published benchmark or quantisation-loss figures; no per-token price on Cohere's own pricing page; production rate limits only through sales; 128K input context; needs vLLM 0.25+ and Cohere's parser library.
- Who is Command A+ best for?
- Enterprises building private, multilingual RAG and agent systems who want a permissively licensed model they can host on two GPUs.
- How much does Command A+ cost for a team or in production?
- Self-hosting needs two H100s or one B200 in W4A4, four H100s in FP8 and eight in BF16. Hosted, 1M RAG answers a month at 8,000 input and 400 output tokens is about $3,000 at OpenRouter's $0.30/$1.50.
- Is Command A+ safe for company data?
- Self-hosted weights keep data in-house under Apache 2.0. Cohere's API platform is SOC 2 Type II with configurable retention and a training opt-out; private deployments run in your VPC, on-premises or in Model Vault, where Cohere says it has no access.
- What are the best Command A+ alternatives?
- The closest alternatives we have reviewed are Mistral Large 4, Mistral Medium 3.5 and GLM-5.3.
Answers checked Oct 22, 2026.